I run onomeo, a small site that gives out free daily credits for 55 models behind one OpenAI-compatible key. Every 8 hours the server sends each model the same prompt, "Reply with the single word: ok", and logs whether it answered and how long it took. This is week one of those logs: Sep 15 11:09 UTC to Sep 21 14:33 UTC, 692 tests.
The big models: 16 of 16 answered every test
GPT-6 Astra, Claude Opus 5, Claude Sonnet 5, Claude Haiku 4.5, Gemini 3.1 Pro, GPT-5.6 Sol, Terra and Luna, Grok 4.6, Kimi K3, GLM-5.3, o3, DeepSeek V4 Pro, Llama 4 Maverick, Mistral Large 3 and Nemotron 3 Ultra joined on Sep 20 and passed all 3 tests since then. Three tests is a small sample; next week's report covers a full week for them. On onomeo they run on the same free daily credits, counted at 1x to 20x depending on the model.
14 free models passed all 19 tests
| Model | Passed | Avg response |
|---|---|---|
| @cf/qwen/qwen2.5-coder-32b-instruct | 19/19 | 0.35 s |
| @cf/meta/llama-4-scout-17b-16e-instruct | 19/19 | 0.36 s |
| ministral-3b-latest | 19/19 | 0.37 s |
| ministral-8b-latest | 19/19 | 0.48 s |
| @cf/openai/gpt-oss-120b | 19/19 | 1.11 s |
| Qwen/Qwen3.8-Flash-Next | 19/19 | 1.61 s |
| mistral-Nemo-Instruct-2407 | 19/19 | 1.93 s |
| codestral-latest | 19/19 | 1.97 s |
| gemini-3-flash-preview | 19/19 | 2.15 s |
| gemini-3.1-flash-lite | 19/19 | 2.24 s |
| deepseek-v4-flash | 19/19 | 2.35 s |
| stepfun/step-3.7-flash:free | 19/19 | 2.68 s |
| Shanghai_AI_Laboratory/Intern-S2-Preview | 19/19 | 3.17 s |
| deepseek-v4.1-flash | 19/19 | 39.85 s |
Cloudflare's hosted models and Mistral's small ones were the fastest. The slowest in that table, deepseek-v4.1-flash, averaged 39.9 s, because it thinks before answering even a one-word prompt.
The flaky ones
| Model | Passed |
|---|---|
| mistralai/mistral-nemotron | 9/19 |
| poolside/laguna-s-2.1:free | 11/19 |
| z-ai/glm-5.3-flash-free | 12/19 |
| tencent/hy3-free | 12/19 |
| nvidia/nemotron-3.5-lightning-30b-a3b | 15/19 |
New and gone
New free models this week: google/gemma-4-26b-a4b-it:free (0/2 so far), inclusionai/ling-3.0-flash-vl:free (1/2 so far), nvidia/nemotron-3-nano-omni-30b-a3b-reasoning:free (0/2 so far).
Gone: ModelScope stopped serving three models. Qwen3-235B-A22B-Thinking-2507 and Qwen3-Coder-30B-A3B-Instruct return "has no provider supported", and GLM-4.7-Flash returns an empty answer.
Caveats
- A one-word prompt shows whether a model is up and how fast it starts. It says nothing about how smart it is.
- Models served through OpenRouter are tested once a day, not every 8 hours.
- A few SiliconFlow failures on Sep 15 were our fault: our account balance had dropped below zero. Fixed on Sep 16.
The live board updates every 8 hours: https://onomeo.com/free-models?w=0922
Disclosure: I run onomeo. This post was drafted with AI help from the site's own test logs.
Top comments (0)